onnx.cpp

#include <google/protobuf/text_format.h>
#include <google/protobuf/io/zero_copy_stream_impl.h>
#include <onnx.pb.h>
#include <iostream>
#include <fstream>
#include <unordered_map>
#include <functional>
#include <array>
#include <utility>
#include <vector>

#include <migraphx/fallthrough.hpp>
#include <migraphx/program.hpp>
#include <migraphx/operators.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/config.hpp>
#include <migraphx/onnx.hpp>

namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {

struct onnx_parser
{
    using attribute_map = std::unordered_map<std::string, onnx::AttributeProto>;
    using node_map      = std::unordered_map<std::string, onnx::NodeProto>;
    using op_func = std::function<instruction_ref(attribute_map, std::vector<instruction_ref>)>;
    node_map nodes;
    std::unordered_map<std::string, instruction_ref> instructions;
    program prog    = program();
    bool is_pytorch = false;

    std::unordered_map<std::string, op_func> ops;
    std::unordered_map<std::string, operation> actv_funcs;

    onnx_parser()
    {
        add_generic_op("MatMul", op::dot{});
        add_generic_op("Relu", op::relu{});
        add_generic_op("Sigmoid", op::sigmoid{});
        add_generic_op("Abs", op::abs{});
        add_generic_op("Exp", op::exp{});
        add_generic_op("Log", op::log{});
        // disable dropout for inference
        add_generic_op("Dropout", op::identity{});
        add_generic_op("Identity", op::identity{});
        add_generic_op("Sin", op::sin{});
        add_generic_op("Cos", op::cos{});
        add_generic_op("Tan", op::tan{});
        add_generic_op("Sinh", op::sinh{});
        add_generic_op("Cosh", op::cosh{});
        add_generic_op("Tanh", op::tanh{});
        add_generic_op("Asin", op::asin{});
        add_generic_op("Acos", op::acos{});
        add_generic_op("Atan", op::atan{});

        add_binary_op("Add", op::add{});
        add_binary_op("Div", op::div{});
        add_binary_op("Mul", op::mul{});
        add_binary_op("Sub", op::sub{});

        add_variadic_op("Sum", op::add{});
        add_variadic_op("Max", op::max{});
        add_variadic_op("Min", op::min{});

        add_mem_op("ImageScaler", &onnx_parser::parse_imagescaler);
        add_mem_op("LeakyRelu", &onnx_parser::parse_leaky_relu);
        add_mem_op("Elu", &onnx_parser::parse_elu);
        add_mem_op("Constant", &onnx_parser::parse_constant);
        add_mem_op("Conv", &onnx_parser::parse_conv);
        add_mem_op("MaxPool", &onnx_parser::parse_pooling);
        add_mem_op("AveragePool", &onnx_parser::parse_pooling);
        add_mem_op("GlobalMaxPool", &onnx_parser::parse_pooling);
        add_mem_op("GlobalAveragePool", &onnx_parser::parse_pooling);
        add_mem_op("Reshape", &onnx_parser::parse_reshape);
        add_mem_op("Flatten", &onnx_parser::parse_flatten);
        add_mem_op("Gemm", &onnx_parser::parse_gemm);
        add_mem_op("BatchNormalization", &onnx_parser::parse_batchnorm);
        add_mem_op("Softmax", &onnx_parser::parse_softmax);
        add_mem_op("Squeeze", &onnx_parser::parse_squeeze);
        add_mem_op("Unsqueeze", &onnx_parser::parse_unsqueeze);
        add_mem_op("Slice", &onnx_parser::parse_slice);
        add_mem_op("Concat", &onnx_parser::parse_concat);
        add_mem_op("Gather", &onnx_parser::parse_gather);
        add_mem_op("Shape", &onnx_parser::parse_shape);
        add_mem_op("ConstantFill", &onnx_parser::parse_constant_fill);
        add_mem_op("Transpose", &onnx_parser::parse_transpose);
        add_mem_op("RNN", &onnx_parser::parse_rnn);

        // init the activation function map
        init_actv_func();
    }

    void init_actv_func()
    {
        actv_funcs.insert(std::make_pair("tanh", op::tanh{}));
        actv_funcs.insert(std::make_pair("relu", op::relu{}));
        actv_funcs.insert(std::make_pair("sigmoid", op::sigmoid{}));
    }

    template <class F>
    void add_op(std::string name, F f)
    {
        ops.emplace(name, f);
    }

    template <class F>
    void add_mem_op(std::string name, F f)
    {
        ops.emplace(name, [=](auto&&... xs) {
            return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
        });
    }

    template <class T>
    void add_binary_op(std::string name, T x)
    {
        ops.emplace(name, [this, x](attribute_map attributes, std::vector<instruction_ref> args) {
            if(args.size() != 2)
                MIGRAPHX_THROW("binary operators should have 2 operands");
            if(contains(attributes, "broadcast"))
            {
                uint64_t broadcasted = parse_value(attributes.at("broadcast")).at<uint64_t>();
                if(broadcasted != 0)
                {
                    uint64_t axis = (contains(attributes, "axis"))
                                        ? parse_value(attributes.at("axis")).at<uint64_t>()
                                        : 0;
                    auto l =
                        prog.add_instruction(op::broadcast{axis, args[0]->get_shape()}, args[1]);
                    return prog.add_instruction(x, args[0], l);
                }
                return prog.add_instruction(x, args);
            }
            else
            {
                return add_broadcastable_binary_op(args[0], args[1], x);
            }
        });
    }

    template <class T>
    instruction_ref add_broadcastable_binary_op(instruction_ref arg0, instruction_ref arg1, T x)
    {
        if(arg0->get_shape() != arg1->get_shape())
        {
            // Example:
            // s0 = (3,2,4,5) and s1 = (2,1,1)
            //
            // In this case we need to broadcast (:,1,1) portion of
            // s1 plus broadcast the 1st dimension of s1
            // giving output_lens = (3,2,4,5)
            //
            // Another example:
            // s0 = (3,2,1,5) and s1 = (2,7,5)
            // In this case we need to broadcast the (:,:,1:,:) axis
            // of s0 plus the 1st dimension of s1 giving
            // output_lens = (3,2,7,5)
            //
            // Get lengths for both arguments
            const std::vector<std::size_t>* s0 = &arg0->get_shape().lens();
            const std::vector<std::size_t>* s1 = &arg1->get_shape().lens();

            // Make sure s0 is the smaller size
            if(s0->size() > s1->size())
                std::swap(s0, s1);

            std::vector<std::size_t> output_lens(*s1);
            auto offset = s1->size() - s0->size();
            std::transform(s0->begin(),
                           s0->end(),
                           s1->begin() + offset,
                           output_lens.begin() + offset,
                           [](auto a, auto b) { return std::max(a, b); });

            auto l0 = prog.add_instruction(op::multibroadcast{output_lens}, arg0);
            auto l1 = prog.add_instruction(op::multibroadcast{output_lens}, arg1);
            return prog.add_instruction(x, l0, l1);
        }
        else
        {
            return prog.add_instruction(x, {arg0, arg1});
        }
    }

    template <class T>
    void add_generic_op(std::string name, T x)
    {
        ops.emplace(name, [this, x](attribute_map, std::vector<instruction_ref> args) {
            return prog.add_instruction(x, args);
        });
    }

    template <class T>
    void add_variadic_op(std::string name, T x)
    {
        ops.emplace(name, [this, x](attribute_map, std::vector<instruction_ref> args) {
            return std::accumulate(std::next(args.begin()),
                                   args.end(),
                                   args.front(),
                                   [this, x](instruction_ref a, instruction_ref b) {
                                       return add_broadcastable_binary_op(a, b, x);
                                   });
        });
    }

    instruction_ref
    parse_softmax(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        auto dims = args.front()->get_shape().lens();
        auto r =
            prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1]), 1, 1}}, args.front());
        auto s = prog.add_instruction(op::softmax{}, r);
        return prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1])}}, s);
    }

    instruction_ref
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::convolution op;
        if(contains(attributes, "pads"))
        {
            if(contains(attributes, "auto_pad"))
            {
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
            }
            std::vector<std::size_t> padding(4);
            copy(attributes["pads"].ints(), padding.begin());
            if(padding.size() != 4)
            {
                MIGRAPHX_THROW("padding should have 4 values");
            }
            if(padding[0] != padding[2] || padding[1] != padding[3])
            {
                MIGRAPHX_THROW("migraphx does not support asymetric padding");
            }
            op.padding[0] = padding[0];
            op.padding[1] = padding[1];
        }
        if(contains(attributes, "strides"))
        {
            copy(attributes["strides"].ints(), op.stride.begin());
        }
        if(contains(attributes, "dilations"))
        {
            copy(attributes["dilations"].ints(), op.dilation.begin());
        }
        if(contains(attributes, "auto_pad"))
        {
            auto s = attributes["auto_pad"].s();
            if(contains(attributes, "pads") and to_upper(s) != "NOTSET")
            {
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
            }

            if(s.find("SAME") != std::string::npos)
            {
                op.padding_mode = op::convolution::same;
            }
        }
        if(contains(attributes, "group"))
        {
            op.group = parse_value(attributes.at("group")).at<int>();
        }
        if(args.size() == 3)
        {
            uint64_t axis = 1;
            auto l1       = prog.add_instruction(op, args[0], args[1]);
            auto l2       = prog.add_instruction(op::broadcast{axis, l1->get_shape()}, args[2]);
            return prog.add_instruction(op::add{}, l1, l2);
        }
        return prog.add_instruction(op, args);
    }

    instruction_ref parse_pooling(const std::string& name,
                                  attribute_map attributes,
                                  std::vector<instruction_ref> args)
    {
        op::pooling op{ends_with(name, "MaxPool") ? "max" : "average"};
        if(starts_with(name, "Global"))
        {
            auto lens  = args.front()->get_shape().lens();
            op.lengths = {lens[2], lens[3]};
        }
        if(contains(attributes, "pads"))
        {
            std::vector<std::size_t> padding(4);
            copy(attributes["pads"].ints(), padding.begin());
            if(padding.size() != 4)
            {
                MIGRAPHX_THROW("padding should have 4 values");
            }
            if(padding[0] != padding[2] || padding[1] != padding[3])
            {
                MIGRAPHX_THROW("migraphx does not support asymetric padding");
            }
            op.padding[0] = padding[0];
            op.padding[1] = padding[1];
        }
        if(contains(attributes, "strides"))
        {
            copy(attributes["strides"].ints(), op.stride.begin());
        }
        if(contains(attributes, "kernel_shape"))
        {
            copy(attributes["kernel_shape"].ints(), op.lengths.begin());
        }
        if(contains(attributes, "auto_pad"))
        {
            auto s = attributes["auto_pad"].s();
            if(to_upper(s) != "NOTSET")
            {
                MIGRAPHX_THROW("auto_pad is not supported for pooling");
            }
        }

        return prog.add_instruction(op, std::move(args));
    }

    instruction_ref
    parse_reshape(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::reshape op;
        if(args.size() == 1)
        {
            literal s = parse_value(attributes.at("shape"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
        }
        if(args.size() == 2)
        {
            literal s = args[1]->get_literal();
            s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
        }
        return prog.add_instruction(op, args[0]);
    }

    instruction_ref
    parse_flatten(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        uint64_t axis = 1;
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
        return prog.add_instruction(op::flatten{axis}, args[0]);
    }

    instruction_ref
    parse_squeeze(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::squeeze op;
        literal s = parse_value(attributes.at("axes"));
        s.visit([&](auto v) { copy(v, std::back_inserter(op.axes)); });
        return prog.add_instruction(op, args[0]);
    }

    instruction_ref
    parse_unsqueeze(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::unsqueeze op;
        literal s = parse_value(attributes.at("axes"));
        s.visit([&](auto v) { copy(v, std::back_inserter(op.axes)); });
        return prog.add_instruction(op, args[0]);
    }

    instruction_ref
    parse_concat(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        std::size_t axis = parse_value(attributes.at("axis")).at<int>();
        op::concat op{axis};
        return prog.add_instruction(op, std::move(args));
    }

    instruction_ref
    parse_gather(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        std::size_t axis = 0;
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
        op::gather op{axis};
        return prog.add_instruction(op, std::move(args));
    }

    instruction_ref
    parse_slice(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::slice op;
        if(contains(attributes, "axes"))
        {
            literal s = parse_value(attributes.at("axes"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.axes)); });
        }
        {
            literal s = parse_value(attributes.at("ends"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.ends)); });
        }
        {
            literal s = parse_value(attributes.at("starts"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.starts)); });
        }
        return prog.add_instruction(op, args[0]);
    }

    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
    {
        literal v = parse_value(attributes.at("value"));
        return prog.add_literal(v);
    }

    instruction_ref
    parse_gemm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        float alpha = 1.0f;
        float beta  = 1.0f;
        bool transa = false;
        bool transb = false;
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        if(contains(attributes, "beta"))
        {
            beta = parse_value(attributes.at("beta")).at<float>();
        }
        if(contains(attributes, "transA"))
        {
            transa = parse_value(attributes.at("transA")).at<bool>();
        }
        if(contains(attributes, "transB"))
        {
            transb = parse_value(attributes.at("transB")).at<bool>();
        }
        std::vector<int64_t> perm = {1, 0};
        auto l1 = (transa) ? prog.add_instruction(op::transpose{perm}, args[0]) : args[0];
        auto l2 = (transb) ? prog.add_instruction(op::transpose{perm}, args[1]) : args[1];
        if(args.size() == 3)
        {
            if(beta != 0.f)
            {
                auto l3 = prog.add_instruction(op::dot{alpha}, l1, l2);
                auto l4 = args[2];
                if(l4->get_shape().scalar()) // ignore args[2] (no C value added to alpha*A*B)
                    return l3;
                if(beta != 1.f)
                {
                    auto beta_val = prog.add_literal(beta);
                    auto l5 = prog.add_instruction(op::scalar{args[2]->get_shape()}, beta_val);
                    l4      = prog.add_instruction(op::mul{}, args[2], l5);
                }
                return add_broadcastable_binary_op(l3, l4, op::add{});
            }
        }
        return prog.add_instruction(op::dot{alpha, beta}, l1, l2);
    }

    instruction_ref
    parse_batchnorm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        float epsilon                                     = 1e-5f;
        float momentum                                    = 0.9f;
        op::batch_norm_inference::bn_infer_mode_t bn_mode = op::batch_norm_inference::spatial;
        bool is_test                                      = false;
        if(contains(attributes, "epsilon"))
        {
            epsilon = parse_value(attributes.at("epsilon")).at<float>();
        }
        if(contains(attributes, "momentum"))
        {
            momentum = parse_value(attributes.at("momentum")).at<float>();
        }
        if(contains(attributes, "is_test"))
        {
            is_test = parse_value(attributes.at("is_test")).at<uint64_t>() > 0;
        }
        if(contains(attributes, "spatial"))
        {
            bn_mode = (parse_value(attributes.at("spatial")).at<uint64_t>() > 0)
                          ? op::batch_norm_inference::spatial
                          : op::batch_norm_inference::per_activation;
        }
        (void)is_test;
        op::batch_norm_inference op{epsilon, momentum, bn_mode};
        return prog.add_instruction(op, std::move(args));
    }

    instruction_ref parse_leaky_relu(const std::string&,
                                     attribute_map attributes,
                                     std::vector<instruction_ref> args)
    {
        float alpha = 0.01; // default alpha val for leaky relu
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        op::leaky_relu op{alpha};
        return prog.add_instruction(op, args.front());
    }

    instruction_ref
    parse_elu(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        float alpha = 1.0; // default alpha val for elu
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        op::elu op{alpha};
        return prog.add_instruction(op, args.front());
    }

    instruction_ref parse_imagescaler(const std::string&,
                                      attribute_map attributes,
                                      std::vector<instruction_ref> args)
    {
        float scale = 1.0;
        std::vector<float> bias{};
        if(contains(attributes, "scale"))
        {
            scale = parse_value(attributes.at("scale")).at<float>();
        }

        if(contains(attributes, "bias"))
        {
            auto&& bias_floats = attributes["bias"].floats();
            bias               = std::vector<float>(bias_floats.begin(), bias_floats.end());
        }
        auto input_shape = args.front()->get_shape();

        auto scale_val = prog.add_literal(scale);
        auto bias_vals = prog.add_literal(
            migraphx::literal{migraphx::shape{migraphx::shape::float_type, {bias.size()}}, bias});

        auto scale_tensor = prog.add_instruction(migraphx::op::scalar{input_shape}, scale_val);
        auto img_scaled   = prog.add_instruction(migraphx::op::mul{}, args.front(), scale_tensor);
        auto bias_bcast = prog.add_instruction(migraphx::op::broadcast{1, input_shape}, bias_vals);
        return prog.add_instruction(migraphx::op::add{}, img_scaled, bias_bcast);
    }

    instruction_ref
    parse_transpose(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        std::vector<int64_t> perm{};
        if(contains(attributes, "perm"))
        {
            auto&& perm_vals = attributes["perm"].ints();
            perm             = std::vector<int64_t>(perm_vals.begin(), perm_vals.end());
        }
        return prog.add_instruction(migraphx::op::transpose{perm}, args.front());
    }

    // Use a literal instruction to replace the shape since, output of
    // shape operator are literals in migraphx
    instruction_ref
    parse_shape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        if(args.size() != 1)
            MIGRAPHX_THROW("Shape: operator should have 1 operand");
        std::vector<std::size_t> arg_shape = args[0]->get_shape().lens();
        std::vector<int64_t> vec_shape(arg_shape.size());
        migraphx::shape s(migraphx::shape::int64_type, {arg_shape.size()});
        std::transform(arg_shape.begin(), arg_shape.end(), vec_shape.begin(), [](auto i) {
            return int64_t(i);
        });
        return prog.add_literal(migraphx::literal{s, vec_shape});
    }

    // Use a literal instruction to replace the constantFill operator. In RNN, input shape
    // and value are fixed, so no need to do the actual computation for the constantFill
    // operator
    instruction_ref parse_constant_fill(const std::string&,
                                        attribute_map attributes,
                                        std::vector<instruction_ref> args)
    {
        int input_as_shape = 0;
        int dtype          = 1;
        float value        = 0.0f;

        if(contains(attributes, "dtype"))
        {
            dtype = parse_value(attributes.at("dtype")).at<int>();
        }
        migraphx::shape::type_t type = get_type(dtype);

        if(contains(attributes, "input_as_shape"))
        {
            input_as_shape = parse_value(attributes.at("input_as_shape")).at<int>();
        }

        if(contains(attributes, "value"))
        {
            value = parse_value(attributes.at("value")).at<float>();
        }

        if(contains(attributes, "extra_shape"))
        {
            MIGRAPHX_THROW("ConstantFill: cannot handle extra shape attribute");
        }

        if(input_as_shape == 1)
        {
            if(args.size() != 1)
            {
                MIGRAPHX_THROW("ConstantFill: need an input argument as output shape");
            }

            if(contains(attributes, "shape"))
            {
                MIGRAPHX_THROW("ConstantFill: cannot set the shape argument and pass in an input "
                               "at the same time");
            }

            migraphx::argument in = args[0]->eval();
            if(in.empty())
            {
                MIGRAPHX_THROW("ConstantFill: cannot handle dynamic shape as input");
            }

            std::vector<std::size_t> dims;
            in.visit([&](auto input) { dims.assign(input.begin(), input.end()); });
            migraphx::shape s(type, dims);
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
        }
        else if(input_as_shape == 0)
        {
            if(!contains(attributes, "shape"))
            {
                MIGRAPHX_THROW("ConstantFill: attribute output shape is needed");
            }

            literal ls = parse_value(attributes.at("shape"));
            std::vector<std::size_t> dims;
            ls.visit([&](auto s) { dims.assign(s.begin(), s.end()); });
            migraphx::shape s{type, dims};
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
        }
        else
        {
            MIGRAPHX_THROW("ConstantFill: wrong value of attribute input_as_shape");
        }
    }

    instruction_ref
    parse_rnn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        migraphx::shape input_shape = args[0]->get_shape();
        migraphx::shape w_shape     = args[1]->get_shape();
        std::size_t hidden_size     = w_shape.lens()[1];

        if(contains(attributes, "hidden_size"))
        {
            hidden_size = parse_value(attributes.at("hidden_size")).at<int>();
        }
        else
        {
            MIGRAPHX_THROW("RNN: hidden size attribute missing");
        }

        std::string activation_func = {"tanh"};
        if(contains(attributes, "activations"))
        {
            activation_func = attributes.at("activations").strings(0);
        }

        if(actv_funcs.count(activation_func) == 0)
        {
            MIGRAPHX_THROW("RNN: activation function " + activation_func + " not supported");
        }

        // Handling of direction to be added later
        std::string direction{"forward"};
        if(contains(attributes, "direction"))
        {
            direction = attributes.at("direction").s();
        }

        op::rnn::rnn_direction_t dirct = op::rnn::forward;
        if(direction == "bidirectional")
        {
            dirct = op::rnn::bidirectional;
        }
        else if(direction == "reverse")
        {
            dirct = op::rnn::reverse;
        }

        // To be added later
        float clip = 0.0;
        if(contains(attributes, "clip"))
        {
            clip = parse_value(attributes.at("clip")).at<float>();
        }

        return prog.add_instruction(op::rnn{hidden_size, actv_funcs[activation_func], dirct, clip},
                                    std::move(args));
    }

    void parse_from(std::istream& is)
    {
        onnx::ModelProto model;
        if(model.ParseFromIstream(&is))
        {
            if(model.has_graph())
            {
                this->parse_graph(model.graph());
            }
        }
        else
        {
            throw std::runtime_error("Failed reading");
        }
    }

    void parse_graph(const onnx::GraphProto& graph)
    {
        nodes = get_nodes(graph);
        std::unordered_map<std::string, onnx::TensorProto> initializer_data;
        for(auto&& f : graph.initializer())
        {
            initializer_data[f.name()] = f;
        }
        for(auto&& input : graph.input())
        {
            const std::string& name = input.name();
            // Does the input have an initializer?
            if(contains(initializer_data, name))
            {
                auto t             = initializer_data[name];
                instructions[name] = prog.add_literal(parse_tensor(t));
            }
            else
            {
                // TODO: Get shape of input parameter
                shape s            = parse_type(input.type());
                instructions[name] = prog.add_parameter(name, s);
            }
        }
        for(auto&& p : nodes)
        {
            this->parse_node(get_name(p.second));
        }
    }

    void parse_node(const std::string& name)
    {
        if(name.empty())
            MIGRAPHX_THROW("Onnx node must have a name");
        if(instructions.count(name) == 0)
        {
            auto&& node = nodes.at(name);
            std::vector<instruction_ref> args;
            for(auto&& input : node.input())
            {
                // For RNN, LSTM, and GRU operators, one of the input arguments
                // is prim::Undefined, and it is ignored by protobuf. We use a
                // hack to ignore this argument for these three operators
                std::string op_type = node.op_type();
                if((op_type == "RNN" || op_type == "LSTM" || op_type == "GRU") &&
                   input.empty() == true)
                {
                    continue;
                }

                if(nodes.count(input) > 0)
                {
                    auto&& iname = get_name(nodes.at(input));
                    assert(name != iname);
                    this->parse_node(iname);
                    args.push_back(instructions.at(iname));
                }
                else
                {
                    args.push_back(instructions.at(input));
                }
            }
            if(ops.count(node.op_type()) == 0)
            {
                instructions[name] = prog.add_instruction(unknown{node.op_type()}, args);
            }
            else
            {
                instructions[name] = ops[node.op_type()](get_attributes(node), args);
            }
        }
    }

    static attribute_map get_attributes(const onnx::NodeProto& node)
    {
        std::unordered_map<std::string, onnx::AttributeProto> result;
        for(auto&& attr : node.attribute())
        {
            result[attr.name()] = attr;
        }
        return result;
    }

    static std::string get_name(const onnx::NodeProto& node)
    {
        if(node.name().empty())
        {
            std::string generated = "migraphx_unnamed_node";
            return std::accumulate(node.output().begin(),
                                   node.output().end(),
                                   generated,
                                   [](auto x, auto y) { return x + "_" + y; });
        }
        return node.name();
    }

    static node_map get_nodes(const onnx::GraphProto& graph)
    {
        std::unordered_map<std::string, onnx::NodeProto> result;
        for(auto&& node : graph.node())
        {
            result[get_name(node)] = node;
            for(auto&& output : node.output())
            {
                result[output] = node;
            }
        }
        return result;
    }

    template <class T>
    static literal from_repeated(shape::type_t t, const T& r)
    {
        std::size_t size = r.size();
        return literal{{t, {size}}, r.begin(), r.end()};
    }

    static literal parse_value(const onnx::AttributeProto& attr)
    {
        switch(attr.type())
        {
        case onnx::AttributeProto::UNDEFINED: return {};
        case onnx::AttributeProto::FLOAT: return literal{attr.f()};
        case onnx::AttributeProto::INT: return literal{attr.i()};
        case onnx::AttributeProto::STRING: return {};
        case onnx::AttributeProto::TENSOR: return parse_tensor(attr.t());
        case onnx::AttributeProto::GRAPH: return {};
        case onnx::AttributeProto::FLOATS: return from_repeated(shape::float_type, attr.floats());
        case onnx::AttributeProto::INTS: return from_repeated(shape::int64_type, attr.ints());
        case onnx::AttributeProto::STRINGS: return {};
        case onnx::AttributeProto::TENSORS: return {};
        case onnx::AttributeProto::GRAPHS: return {};
        }
        MIGRAPHX_THROW("Invalid attribute type");
    }

    static literal parse_tensor(const onnx::TensorProto& t)
    {
        std::vector<std::size_t> dims(t.dims().begin(), t.dims().end());
        // in case of scalar constants in onnx file, use dims=1 to fill initializer data
        if(dims.empty())
        {
            dims = {1};
        }
        if(t.has_raw_data())
        {
            const std::string& s = t.raw_data();
            switch(t.data_type())
            {
            case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
            case onnx::TensorProto::FLOAT: return literal{{shape::float_type, dims}, s.data()};
            case onnx::TensorProto::UINT8: throw std::runtime_error("");
            case onnx::TensorProto::INT8: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::UINT16: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::INT16: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::INT32: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::INT64: return literal{{shape::int64_type, dims}, s.data()};
            case onnx::TensorProto::STRING: throw std::runtime_error("");
            case onnx::TensorProto::BOOL: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::FLOAT16: return literal{{shape::half_type, dims}, s.data()};
            case onnx::TensorProto::DOUBLE: return literal{{shape::double_type, dims}, s.data()};
            case onnx::TensorProto::UINT32: throw std::runtime_error("");
            case onnx::TensorProto::UINT64: throw std::runtime_error("");
            case onnx::TensorProto::COMPLEX64: throw std::runtime_error("");
            case onnx::TensorProto::COMPLEX128: throw std::runtime_error("");
            }
            MIGRAPHX_THROW("Invalid tensor type");
        }
        switch(t.data_type())
        {
        case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
        case onnx::TensorProto::FLOAT:
            return literal{{shape::float_type, dims}, t.float_data().begin(), t.float_data().end()};
        case onnx::TensorProto::UINT8: throw std::runtime_error("");
        case onnx::TensorProto::INT8:
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
        case onnx::TensorProto::UINT16:
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
        case onnx::TensorProto::INT16:
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
        case onnx::TensorProto::INT32:
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
        case onnx::TensorProto::INT64:
            return literal{{shape::int64_type, dims}, t.int64_data().begin(), t.int64_data().end()};
        case onnx::TensorProto::STRING: throw std::runtime_error("");
        case onnx::TensorProto::BOOL:
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
        case onnx::TensorProto::FLOAT16:
            return literal{{shape::half_type, dims}, t.float_data().begin(), t.float_data().end()};
        case onnx::TensorProto::DOUBLE:
            return literal{
                {shape::double_type, dims}, t.double_data().begin(), t.double_data().end()};
        case onnx::TensorProto::UINT32: throw std::runtime_error("");
        case onnx::TensorProto::UINT64: throw std::runtime_error("");
        case onnx::TensorProto::COMPLEX64: throw std::runtime_error("");
        case onnx::TensorProto::COMPLEX128: throw std::runtime_error("");
        }
        MIGRAPHX_THROW("Invalid tensor type");
    }

    static shape parse_type(const onnx::TypeProto& t)
    {
        shape::type_t shape_type{};
        switch(t.tensor_type().elem_type())
        {
        case onnx::TensorProto::UNDEFINED:
            break; // throw std::runtime_error("Unsupported type UNDEFINED");
        case onnx::TensorProto::FLOAT: shape_type = shape::float_type; break;
        case onnx::TensorProto::UINT8:
            break; // throw std::runtime_error("Unsupported type UINT8");
        case onnx::TensorProto::INT8: shape_type = shape::int8_type; break;
        case onnx::TensorProto::UINT16: shape_type = shape::uint16_type; break;
        case onnx::TensorProto::INT16: shape_type = shape::int16_type; break;
        case onnx::TensorProto::INT32: shape_type = shape::int32_type; break;
        case onnx::TensorProto::INT64: shape_type = shape::int64_type; break;
        case onnx::TensorProto::STRING:
            break; // throw std::runtime_error("Unsupported type STRING");
        case onnx::TensorProto::BOOL:
            break; // throw std::runtime_error("Unsupported type BOOL");
        case onnx::TensorProto::FLOAT16: shape_type = shape::half_type; break;
        case onnx::TensorProto::DOUBLE: shape_type = shape::double_type; break;
        case onnx::TensorProto::UINT32: shape_type = shape::uint32_type; break;
        case onnx::TensorProto::UINT64: shape_type = shape::uint64_type; break;
        case onnx::TensorProto::COMPLEX64:
            break; // throw std::runtime_error("Unsupported type COMPLEX64");
        case onnx::TensorProto::COMPLEX128:
            break; // throw std::runtime_error("Unsupported type COMPLEX128");
        }
        std::vector<std::size_t> dims;
        auto&& tensor_dims = t.tensor_type().shape().dim();
        std::transform(tensor_dims.begin(),
                       tensor_dims.end(),
                       std::back_inserter(dims),
                       [](auto&& d) -> std::size_t {
                           if(not d.has_dim_value())
                           {
                               long default_batch_size = 1; // FIXME
                               return default_batch_size;
                           }
                           return d.dim_value();
                       });
        return {shape_type, dims};
    }

    shape::type_t get_type(int dtype)
    {
        switch(dtype)
        {
        case 1: return shape::float_type;
        case 2: return shape::uint8_type;
        case 3: return shape::int8_type;
        case 4: return shape::uint16_type;
        case 5: return shape::int16_type;
        case 6: return shape::int32_type;
        case 7: return shape::int64_type;
        case 10: return shape::half_type;
        case 11: return shape::double_type;
        case 12: return shape::uint32_type;
        case 13: return shape::uint64_type;
        default:
        {
            MIGRAPHX_THROW("Prototensor data type " + std::to_string(dtype) + " not supported");
        }
        }
    }
};

program parse_onnx(const std::string& name)
{
    std::fstream input(name.c_str(), std::ios::in | std::ios::binary);
    onnx_parser parser;
#ifndef NDEBUG
    // Log the program when it can't be parsed
    try
    {
        parser.parse_from(input);
    }
    catch(...)
    {
        std::cerr << parser.prog << std::endl;
        throw;
    }
#else
    parser.parse_from(input);
#endif
    return std::move(parser.prog);
}

} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx